The report covers possible machine-learning interventions in 13 domains, from electricity systems to farms and forests to climate prediction. Within each domain, it breaks out the contributions for various subdisciplines within machine learning, including computer vision, natural-language processing, and reinforcement learning.
Recommendations are also divided into three categories: “high leverage” for problems well suited to machine learning where such interventions may have an especially great impact; “long-term” for solutions that won’t have payoffs until 2040; and “high risk” for pursuits that have less certain outcomes, either because the technology isn’t mature or because not enough is known to assess the consequences. Many of the recommendations also summarize existing efforts that are already happening but not yet at scale.
The report’s compilation was led by David Rolnick, a postdoctoral fellow at the University of Pennsylvania, and advised by several high-profile figures, including Andrew Ng, the cofounder of Google Brain and a leading AI entrepreneur and educator; Demis Hassabis, the founder and CEO of DeepMind; Jennifer Chayes, the managing director of Microsoft Research; and Yoshua Bengio, who recently won the Turing Award for his contributions to the field. While the researchers offer a very comprehensive list of some of the major areas where machine learning can contribute, they also note that it is not a silver bullet. Ultimately, policy will be the main driver for effective large-scale climate action.
Here are just 10 of the “high leverage” recommendations from the report. Read the full version of it here.
If we’re going to rely on more renewable energy sources, utilities will need better ways of predicting how much energy is needed, in real time and over the long term. Algorithms already exist that can forecast energy demand, but they could be improved by taking into account finer local weather and climate patterns or household behavior. Efforts to make the algorithms more explainable could also help utility operators interpret their outputs and use them in scheduling when to bring renewable sources online.
Scientists need to develop materials that store, harvest, and use energy more efficiently, but the process of discovering new materials is typically slow and imprecise. Machine learning can accelerate things by finding, designing, and evaluating new chemical structures with the desired properties. This could, for example, help create solar fuels, which can store energy from sunlight, or identify more efficient carbon dioxide absorbents or structural materials that take a lot less carbon to create. The latter materials could one day replace steel and cement—the production of which accounts for nearly 10% of all global greenhouse-gas emissions.
Shipping goods around the world is a complex and often highly inefficient process that involves the interplay of different shipment sizes, different types of transportation, and a changing web of origins and destinations. Machine learning could help find ways to bundle together as many shipments as possible and minimize the total number of trips. Such a system would also be more resilient to transportation disruptions.
Electric vehicles, a key strategy for decarbonizing transportation, face several adoption challenges where machine learning could help. Algorithms can improve battery energy management to increase the mileage of each charge and reduce “range anxiety,” for example. They can also model and predict aggregate charging behavior to help grid operators meet and manage their load.
Intelligent control systems can dramatically reduce a building’s energy consumption by taking weather forecasts, building occupancy, and other environmental conditions into account to adjust the heating, cooling, ventilation, and lighting needs in an indoor space. A smart building could also communicate directly with the grid to reduce how much power it is using if there’s a scarcity of low-carbon electricity supply at any given time.
Many regions of the world have little to no data on their energy consumption and greenhouse-gas emissions, which can be a major obstacle to designing and implementing effective mitigation strategies. Computer vision techniques can extract building footprints and characteristics from satellite imagery to feed machine-learning algorithms that can estimate city-level energy consumption. The same techniques could also identify which buildings should be retrofitted to maximize their efficiency.
In the same way that machine learning can optimize shipping routes, it can also minimize inefficiencies and carbon emissions in the supply chains of the food, fashion, and consumer goods industries. Better predictions of supply and demand should significantly reduce production and transportation waste, while targeted recommendations for low-carbon products could encourage more environmentally friendly consumption.
Much of modern-day agriculture is dominated by monoculture, the practice of producing a single crop on a large swath of land. This approach makes it easier for farmers to manage their fields with tractors and other basic automated tools, but it also strips the soil of nutrients and reduces its productivity. As a result, many farmers rely heavily on nitrogen-based fertilizers, which can convert into nitrous oxide, a greenhouse gas 300 times more potent than carbon dioxide. Robots run on machine-learning software could help farmers manage a mix of crops more effectively at scale, while algorithms could help farmers predict what crops to plant when, regenerating the health of their land and reducing the need for fertilizers.
Deforestation contributes to roughly 10% of global greenhouse-gas emissions, but tracking and preventing it is usually a tedious manual process that takes place on the ground. Satellite imagery and computer vision can automatically analyze the loss of tree cover at a much greater scale, and sensors on the ground, combined with algorithms for detecting chainsaw sounds, can help local law enforcement stop illegal activity.
Techniques that advertisers have successfully used to target consumers can be used to help us behave in more environmentally aware ways. Consumers could receive tailored interventions to promote their enrollment in energy saving programs, for example.